Jupyter Agent: 10K-100K Executable Reasoning Traces from Kaggle

10,000 to 100,000 records of Kaggle notebooks transformed into executable reasoning traces and question-answer pairs by jupyter-agent in 2025. The collection features natural language questions grounded in specific datasets and verified answers produced by running the original code.

Use Cases

• Training autonomous agents to generate executable code using reasoning traces
• Benchmarking question-answering models on dataset-grounded QA pairs
• Fine-tuning code generation models for Polars and Dask library syntax

Strengths

• 10,000 to 100,000 records
• Executable reasoning traces verified by notebook runs
• Educational quality scoring used for filtering

Limitations

• Machine-generated annotations may introduce synthetic noise or logic errors
• Scope is restricted to notebooks relevant to data analysis tasks

Provenance

Source

Kaggle

Collection Method

Multi-stage pipeline involving de-duplication, quality scoring, and synthetic QA generation

Freshness

Last updated September 2025